Machine learning: Ensemble trees in a time series classification problem.

Dear all,

I have an issue with the ensemble trees algorithms.

I am woking in a time series forecasting problem. As these algos are dividing the training set randomly to create many trees it will train a tree based on future values is not consistant. Indeed, it performs very badly out of sample.

Do you have any idea how I could handle that ? Any parameter available ?

Thank you very much in advance and have a nice weekend.

Pierre

Answer 1

[Bernhard Suhm]

How are you encoding the time dependency in your predictors? If you add predictors that look back fixed number of elements in your time series, it shouldn't matter that the ensemble algorithms randomly selects a subset of samples to grow the tree. Each sample will consist of the time series value and its history in pre-determined intervals. Or if I'm missing something, provide more detail about how you're constructing your predictors.